Protein structure generation via folding diffusion.

Area of Science:

• Computational biology
• Structural biology
• Biophysics

Background:

• Generating novel, physically foldable protein structures computationally is crucial for biological discovery and therapeutic development.
• Current neural network approaches struggle to produce diverse and novel protein structures directly.
• Understanding protein folding mechanisms is key to designing new proteins.

Purpose of the Study:

• To develop a diffusion-based generative model for creating novel protein backbone structures.
• To leverage principles of natural protein folding in a computational generation process.
• To overcome limitations of existing neural network models in generating diverse protein structures.

Main Methods:

• Representing protein backbones as sequences of angles describing atomic orientations.
• Employing a denoising diffusion probabilistic model (DDPM) trained on this representation.
• Utilizing a simple transformer architecture within the DDPM.
• Generating structures by denoising from a random state to a stable conformation.

Main Results:

• The model unconditionally generates highly realistic protein structures.
• Generated structures exhibit complexity and patterns similar to naturally occurring proteins.
• The angle-based representation simplifies the model by avoiding complex equivariant networks.

Conclusions:

• Diffusion models offer a promising approach for de novo protein structure generation.
• The developed method successfully generates diverse and realistic protein backbones.
• The open-source release facilitates further research in computational protein design.